New Framework Generalizes Distribution-Free Semi-Supervised Learning

Traditional semi-supervised learning (SSL) methods often rely on specific distributional assumptions, which can limit their effectiveness when these assumptions are not met. While PNU learning offered a distribution-free alternative, it was restricted to binary classification and its variance optimality was not fully understood. A new generalized framework addresses these limitations by constructing unbiased risk estimators through linear combinations of component risks. This approach not only subsumes PNU learning but also extends its applicability to multiclass classification problems. The framework demonstrates that it can achieve lower variance than PNU learning, particularly in scenarios involving asymmetric losses. Theoretical insights also establish a direct link between this variance reduction and improved learning performance, with two new practical SSL methods empirically matching or outperforming existing approaches on both binary and multiclass benchmarks.